End to End Learning

End-to-end learning aims to train entire systems, from raw input to final output, without manually designing intermediate steps. Current research focuses on applying this approach to diverse problems, including robotics (grasping, navigation, control), computer vision (object recognition, pose estimation, segmentation), and signal processing (communications, speech recognition), often employing neural networks, particularly recurrent and transformer architectures. This paradigm offers the potential for improved performance and generalization by optimizing all components jointly, leading to more efficient and robust systems across various scientific and engineering domains.

Papers